Talk:Graph Neural Networks

Very nicely written and well explained. I personally liked the mention of the limitations - "Problems with Graph Neural Networks". I have some minor suggestions:

Maybe you can elaborate a bit more on the adjacency list, where you use the same example that you used for the adjacency matrix also to describe adjacency list and how it differs from the matrix.

It was a bit unclear: "There are broadly three kinds of tasks on graphs,"

You can also explain the figure of Zach's Karate Club with a small text description. I found it a bit difficult to connect the image with the text on node-level prediction.

It was slightly confusing: "*Note: A pooling layer is generally present if we need a graph-level representation/embedding. This is useful when we want to do graph-level tasks." In what cases is the pooling layer not available?

I like the discussion on permutation invariance and equivariance. I think it is a very important property of GNNs.

Can you also give one-line simplified motivation/intuition behind GCN and Graph attention Networks?

Thank you so much for the review. I have made changes accordingly.

> Maybe you can elaborate a bit more on the adjacency list, where you use the same example that you used for the adjacency matrix also to describe adjacency list and how it differs from the matrix.

Added a short explanation on how it would look like.

> It was a bit unclear: "There are broadly three kinds of tasks on graphs,"

Changed this.

> You can also explain the figure of Zach's Karate Club with a small text description. I found it a bit difficult to connect the image with the text on node-level prediction.

It was initially in the node classification description. Moved it to the image caption.

> It was slightly confusing: "*Note: A pooling layer is generally present if we need a graph-level representation/embedding. This is useful when we want to do graph-level tasks." In what cases is the pooling layer not available?

I just removed it. I think the complication wasn't needed.

The article is quite interesting and is mostly easy to grasp for a beginner. Some suggestions from my end.

- The motivation for both the message passing and pooling layer is not clear. Maybe adding one or two lines each for both of them would help.

- Can you go over different pooling schemes since that doesn't seem to have been covered? Also, the note about pooling makes it ambiguous. See if it can be explained better.

- Can you maybe show what an adjacency list looks like? You can elaborate a bit more in the description. Also, highlight what of adjacency matrix/list is used during message passing.

- The first limitation of the WL-test is hard to understand for someone without a graph background. Can you maybe motivate it with the example that you have shown and then go on to the theoretical bits?

- Can you highlight some python implementations that could be useful for anyone starting out with graph neural networks?

Thank you so much for the feedback. I have made changes accordingly.

> - The motivation for both the message passing and pooling layer is not clear. Maybe adding one or two lines each for both of them would help.

Added a short line when the two are introduced.

> - Can you go over different pooling schemes since that doesn't seem to have been covered? Also, the note about pooling makes it ambiguous. See if it can be explained better.

I have added commonly used pooling strategies.

> - Can you maybe show what an adjacency list looks like? You can elaborate a bit more in the description. Also, highlight what of adjacency matrix/list is used during message passing.

Added this for the same graph in the figure.

> - The first limitation of the WL-test is hard to understand for someone without a graph background. Can you maybe motivate it with the example that you have shown and then go on to the theoretical bits?

Changed the order of explanation.

> - Can you highlight some python implementations that could be useful for anyone starting out with graph neural networks?

Added a section on this.